Increasing demand on public healthcare services due to the aging population has become a major problem, driving a need to build proactive healthcare systems. In such systems, activity recognition is an indispensable component. Quantification of daily physical activity is a key determinant in evaluation of the quality of life of subjects with limited mobility, such as elderly persons . By analyzing human’s activity, much useful information about human’s health condition could be extracted.
The quantitative assessment of daily activity in humans requires an objective and reliable technique that can be used under conditions of daily living. Complex sensors such as cameras in computer vision have been extensively used for recognizing activities. Computer vision sensing for tracking and recognizing activity often works in the laboratory but fails in real home settings due to clutter, variable lighting, and highly varied activities that take place in natural environments.  Complexity of dealing with changes in the scene, such as lighting, multiple people and clutter offers additional challenges. Finally, because sensors such as microphones and cameras record biometric information, they can also be perceived as invasive (i.e., violating privacy) by some people.
A range of body-fixed sensors including electromechanical switches (attached to the heel to identify timing of heel strike in gait), goniometers (to measure joint angles between body segments), accelerometers (to measure motion of body segments), gyroscopes (to measure orientation of body segments), pedometers and actometers have been used to measure aspects of human movement in free-living subjects as shown in Fig 3. Of these, accelerometers are becoming widely accepted as a useful tool for the assessment of human motion in clinical settings and free-living environments. Accelerometers offer a number of desirable features in monitoring of human movement. Firstly, they respond to both frequency and intensity of movement, and so are superior to actometers or pedometers, which are attenuated by impact or tilt. Secondly, some types of accelerometers can be used to measure tilt as well as body movement, making them superior to motion sensors that have no ability to measure static characteristics. Thirdly, enhancements in microelectromechanical systems (MEMS) technology have made possible the manufacture of miniaturized, low cost accelerometers. These instruments also demonstrate a high degree of reliability in measurement, with little variation over time. This has enabled the development of small, lightweight, portable systems that can be worn by a free-living subject without impeding movement. Systems can be designed that are suitable for monitoring in the patient’s normal environment over extended periods.

Some previous studies, incorporating the use of accelerometers, have been reported to identify the type of activity, but their methods are cumbersome because they use two or more different sites of attachments to the body and cable connections, reducing their applicability for long-term monitoring of physical activity: in fact they interfere with normal activities. Comparatively, a very small number of studies have investigated the use of a single accelerometer mounted at waist, sternum or back.
In their works, a large number of recognition methods have been investigated. Many studies incorporated the idea of simple heuristic classifiers for the classification of possible motions and postures. While other employed more generic and automatic methods from machine learning including decision trees, nearest neighbor and Bayes, support vector machines, neural networks, Gaussian mixture models and Markov chains.
Thus, existing literature on movement classification using accelerometry data widely varies in approach, intention and outcome. Individual researchers had investigated their own set of movements using their own devices and their own data collection method and have applied a wide variety of algorithms and methods. Consequently, it is difficult to make significant comparisons or draw meaningful conclusions from the existing literature beyond noting that accelerometer shows promise in human activity monitoring and recognition.

The aim of our research is to build a new framework of human activity recognition (HAR) system which provides an effective and robust approach for the recognition and classification of different movements and postures from data obtained using a single, chest mounted triaxial accelerometer; in particular aiming to improve the flexibility and generality of the monitoring system by working along two dimensions i.e.

• Signal Modeling: We propose to model the acceleration signal using different modeling techniques e.g. AR-Modeling, ARMA Modeling, Kalman Filters etc, and use the coefficients of these models as features for classification of human motions and postures.
• Senor Position Independence: In real life, users would not be willing to have accelerometers fixed on their bodies. Instead these motion sensors would be carried along as a part of hand held devices e.g. mobile phones. Since users can carry these hand held devices in their pockets, hands or bags, sensor orientation and position cannot be fixed at one point. Our aim is to develop a recognition method that would provide acceptable accuracy rates independent of the sensor’s position on subject’s body.

Figure 4 explains the overall technique as a block diagram. The main components of our proposed system are explained below.

• Augmented AR Coefficients: We propose a general, but new framework of human activity recognition (HAR) system using a novel set of features derived from a single triaxial accelerometer and artificial neural nets (ANNs). We use the autoregressive (AR) modeling coefficients of activity signals as key features, but augment them with Signal Magnitude Area (SMA) and Tilt Angle (TA) to distinguish dynamic activity from static, and a general ANNs for recognition.  
• Wavelet Coefficients: The use of wavelet coefficients for human activity recognition using accelerometer had been investigated by different researchers. We proposed to use the entropy measurement of the wavelet coefficients to come up with the best decomposition level for the wavelet packet decomposition of the acceleration signal and then use these coefficients as a bench mark to analyze the accuracy of our proposed technique i.e. Augmented AR Coefficients.

• Statistical Features: Statistical measurements like mean, std, variance, energy, spectral entropy, correlation among different axes etc had been used in past researches as features for classifying human motions and postures into different classes. Here we propose to use these measurements as a single feature vector, computed over different window sizes for activity clasification and using these results as another benchmark for analyzing the preformance of our proposed technique i.e. Augmented AR Coefficients.